How to Handle Sensitive Customer Conversations Without AI

Keep AI out of conversations where a wrong word causes real harm or the customer needs a person: bereavement, serious complaints and legal threats, money trouble, health or vulnerability disclosures, fraud or scam victims, injuries or major losses, and anyone who asks for a human. Route those to a named person quickly, with the context already gathered.

The test is simple: what does a mistake cost, and who carries it? An assistant that gives a slightly clumsy answer about a renewal date costs nothing. One that tells a widow the policy "will be updated automatically", or offers a payment holiday nobody can promise, costs trust, money and sometimes a complaint to a regulator. AI is fast and consistent; these conversations need judgement, accountability and a name.

Follow me on Instagram@sagnikteaches

Seven kinds of conversation to keep with people

Each of these has a different reason for keeping AI out, and knowing the reason helps staff spot new cases that don't fit the list.

Connect on LinkedInSagnik Bhattacharya

1. Bereavement. The customer is often dealing with several organisations at once. Automated sympathy followed by a form link is the experience people remember and tell others about.
2. Complaints that mention lawyers, regulators or going public. The first reply can shape a dispute. An assistant can admit fault, argue, or promise a refund, and each of those is hard to walk back.
3. Money trouble. "I've lost my job", "I can't make the payment". The right answer depends on the customer's situation and on decisions other people (lenders, insurers) make. Any confident answer from a bot is a promise nobody authorised.
4. Health, disability and other vulnerability. Disclosures change how you should communicate and sometimes what you can offer. They also involve sensitive personal data you shouldn't be pushing through extra systems.
5. Fraud and scams. A customer who has just been scammed is anxious and time matters. They need a person who can act, and an assistant that can be impersonated or manipulated is the wrong first contact.
6. Injury, fire, flood or major loss. The situation is urgent and emotional, and the practical next steps depend on details a person needs to hear.
7. Anyone who asks for a person. Keeping someone in a loop with a bot after they've asked to leave it turns a routine query into a complaint.

The broader question of which business tasks should stay human, beyond customer conversations, is covered in when not to use AI in your business.

Subscribe on YouTube@codingliquids

A routing table that catches them early

Routing works from signals: the words people use and the situations they describe. Most chat and help desk tools let you route by keywords, topics or detected intent; whatever your tool calls it, the table below is the content you give it. A filled-in version for an illustrative insurance broker handling home, motor and small-business policies:

Two rules make the table work. Route on the first signal, even mid-conversation: if the customer mentions a funeral in their third message, the assistant stops answering and hands over. And set the tool to hand over whenever it isn't confident what a message is about, rather than guessing. The general principles are set out in the tutorial on chatbot handover timing.

The handover message: written by a person, sent by the bot

The one thing the assistant does say in these conversations is the handover message, and it should be fixed text a person wrote in advance, not something generated on the spot. Generated sympathy is where the damage happens. An illustrative default from an assistant left to improvise:

I'm so sorry for your loss. I understand this must be an
incredibly difficult time for you. I've passed your message to
our team, who will be in touch shortly. In the meantime, you
may find our guide to updating your policy details helpful!

Four problems: sympathy performed by software, "shortly" promising nothing, a self-service link at the worst possible moment, and an exclamation mark. The broker's pre-written version, which the tool sends word for word when the bereavement route fires:

Thank you for letting us know. I'm an automated assistant, so I've passed your message straight to [account manager's first name], who looks after your policy. They will call you today before 4pm on the number we hold for you. If you'd rather we emailed, or another time suits you better, just reply here.

It is honest about being automated, names a person, gives a time the business will actually meet, and hands control back to the customer. Write one message per route in the table; seven short templates cover almost everything.

The note that goes with every handover

Nothing makes a sensitive conversation worse than asking the customer to tell their story twice. The handover should carry a note so the person picking it up starts informed. A filled-in example from the broker, for the bereavement route:

The "what the assistant said" line matters more than it looks. If the bot gave any information before the route fired, the person needs to know exactly what, so they can correct it if it was wrong.

Five moves for the conversation itself

Handling difficult customer conversations well is a skill, but it follows a structure staff can learn. Five moves, in order:

1. Acknowledge once, plainly. "I'm sorry about your husband" is enough. Repeated sympathy starts to sound scripted, which is the thing you took the bot out to avoid.
2. Establish what they need. Ask what prompted the contact today. People in hard situations often lead with the wrong question because it's the one they could put into words.
3. Set out the options, including doing nothing yet. "Nothing needs to change today" is often the most useful sentence in the call.
4. Give them an owner. Your name, your direct line, and that you will handle it from here.
5. Agree a date. When you will next be in touch and what you will have done by then. Then do it by that date.

Written down, these look obvious. Their value is that they give a nervous member of staff something to hold on to, and they are the opposite of what an assistant does by default: repeated sympathy, a generic answer, and "the team".

A bereavement call at a financial planning firm

An illustration of the five moves in one call. The adult son of a long-standing client of a financial planning firm emails the firm's website assistant: "My mother died last week. She had investments with you. What do we need to do?" The bereavement route fires, the pre-written message goes out, and the client's planner calls that afternoon.

• Acknowledge: the planner knew his mother for eleven years and says so, briefly.
• Establish: he is one of two executors and is worried about deadlines, because someone told him accounts get frozen.
• Options: before discussing any details of the accounts, the firm needs to confirm who has authority to act, so the planner explains what documents that involves. Nothing is lost by taking a few weeks to gather them, and the planner explains which parts genuinely can't wait (none, in this case).
• Owner: the planner gives his direct line and says he will deal with the other executor too, so they don't get conflicting messages.
• Date: he will send a one-page list of what's needed by Friday, and call again in two weeks.

An assistant in that conversation would have faced two traps: answering "what do we need to do?" with a generic bereavement checklist from the internet, and discussing the mother's investments with someone whose authority hadn't been confirmed. The second is a confidentiality breach, whatever the intention. The firm's whole AI involvement in this case was the routing and the fixed message, which is as it should be.

Money trouble: why a mortgage adviser answers in person

A customer messages an illustrative mortgage adviser's website late on a Sunday: "I've been made redundant and I don't think I can make next month's payment. Can I take a payment holiday?" An assistant trained on the adviser's web pages might answer with what it has read: that payment holidays exist and how they work. But the lender decides, not the adviser, and a customer who hears "yes, you can" from the adviser's own website may stop paying before anything is agreed.

The routed version: the pre-written message says an adviser will call on Monday morning and, until then, the most useful step is not to ignore the lender's letters. On Monday, the adviser explains that the lender is the one to contact and why contacting them early usually opens more options than waiting, talks through what those options can include without promising any, and points the customer to free, independent debt advice services. The adviser also checks whether any protection policy arranged through the firm might pay out after redundancy. That last check is exactly the kind of joined-up thinking an assistant answering one question at a time won't do. How advisers draw the line between automation and advice more generally is in what mortgage brokers can automate with AI and what stays advice.

Complaints that mention lawyers or regulators

When a customer writes "I'll be taking this to my lawyer", the first reply has one job: acknowledge the complaint and hand it to the person responsible, without arguing the substance. A person writes it. The illustrative law firm's complaints owner uses this structure within the same working day:

• Confirm the complaint has been received and logged, with a reference.
• Name the person handling it and their direct contact details.
• State the timescale from the firm's own complaints procedure, and when the next update will come.
• Invite anything further they want considered.
• Say nothing yet about fault, refunds or what went wrong.

An assistant replying to the same message tends to do the opposite: apologise for "the error" (an admission), explain why the firm acted correctly (an argument), or offer "to look into a refund" (a promise). The line between complaints AI can help with and those it makes worse is drawn in more detail in whether AI can handle customer complaints.

Testing the routing with ten awkward messages

Routing rules look complete on paper and leak in practice, because customers don't use your keywords. Before switching the rules on, and every quarter after, send ten test messages written the way stressed people actually write. The broker's first test, illustrative results:

Eight of ten correct. The fixes: add "passed on" and other phrasings, and add a rule that any message expressing that the customer can't cope or can't pay, however worded, hands over; most tools that detect intent or sentiment can do this better than a keyword list. Then keep the weekly check running, because live customers will find wordings no test anticipated. The weekly sampling routine for AI support replies gives missed handovers their own place in the sample.

What these conversations cost in staff time, and who carries them

Keeping these conversations human has a cost, and it's worth sizing so it can be staffed properly rather than squeezed in. The illustrative broker's assistant handles about 1,800 conversations a month. In testing, roughly 6% hit a sensitive route: about 108 a month. At an average of 20 minutes of staff time each, including the note and follow-up, that is 36 hours a month, or about 9 hours a week, split between two named handlers and a director for fraud and complaints. The assistant keeps the other 94%.

The people taking these conversations need looking after too. Rotate the routes so one person isn't handling every bereavement and every angry complaint; allow ten minutes after a hard call before the next task; make clear that staff can end a call that becomes abusive and pass it on. The templates and five moves exist to support staff, not to script them: a person who departs from the template because the customer needs something different is doing the job right.

AI can still help around these conversations, out of the customer's sight: spotting the signals and routing, reminding a handler that a promised call is due, or generating practice scenarios with invented details for training. What it shouldn't do in these seven situations is write the words the customer reads or decide what the customer is offered.

Sensitive conversations and AI: more questions

Can AI draft the reply if a person checks it before it's sent?

For these conversations, it's better if the person writes it. A reviewed AI draft tends to keep the draft's phrasing: stock sympathy, vague timings, a helpful link that shouldn't be there. Pre-written templates that a person adapts give you consistency without that risk. Keep AI for routing and internal admin around the conversation, not the words the customer reads.

What if a customer in a sensitive situation prefers the chatbot?

Some people do, especially out of hours or when a subject is embarrassing to say aloud. Offer the choice: the assistant can hand over and tell them when a person will reply, while making clear they can keep using it for simple things like documents or dates. What it shouldn't do is give answers about money, cover, health or complaints in that state.

Are there legal rules about vulnerable customers?

Some sectors have them. Many financial regulators, for example, publish expectations about identifying and supporting customers in vulnerable circumstances, and health details usually carry extra protection under data-protection law. Check what applies to your sector and ask your compliance or data-protection adviser; the steps here are about practical handling, not legal obligations.

Further reads

• How to Write Call Scripts and Escalation Rules for an AI Receptionist — Escalation rules for AI answering the phone, not just chat.
• Chatbot Guardrails: Stop AI Promising What You Don't Offer — Guardrails for the conversations the assistant does keep.
• AI Chatbot Disclosure: What to Tell Customers at the Start of a Chat — What to tell customers about the assistant at the start.
• How Insurance Brokers Use AI to Handle Claims Enquiries — Where AI helps with claims enquiries and where people take over.
• How to Write Reply Templates That Keep AI Replies On-Script — Reply templates for the routine conversations AI keeps.
• Who Is Liable When Your AI Chatbot Gets It Wrong? — Who carries the risk when a bot answers what it shouldn't.
• How Spas Can Upsell Add-On Treatments With AI Without Being Pushy — A pairing matrix, booking-screen add-on settings, a pre-visit message prompt and therapist wording that sells add-ons without pressuring anyone.
• How Personal Trainers Can Use ChatGPT Safely for Client Programmes — What to type and what to leave out, a prompt that builds in each client's limits, a five-minute check before sending, and where ChatGPT must stop.
• How Funeral Directors Can Use AI for Admin With Care — Where AI can take funeral admin off your desk, where it must never speak for you, and the read-back routine that stops a misspelt name reaching print.
• Which Pharmacy Tasks Should Never Be Handed to AI? — The pharmacy tasks that must stay with a pharmacist, why each one fails with AI, and the admin work around them that AI can safely support.
• Can Customers Tell When a Reply Was Written by AI? — What customers actually notice in AI-drafted replies, what research says about spotting AI text, and a rule for which replies AI should draft at all.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.